ATLAS Jet Reconstruction, Calibration, and Tagging

Transcription

1 ALAS Jet Reconstruction, Calibration, and agging Steven Schramm On behalf of the ALAS Collaboration ICNFP 7 Crete, Greece August, 7 Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 /

2 Introduction Left: EXO-6- Right: PERF-4-7 Introduction to s Jets are a tool to represent hadronic showers in a detector ALAS primarily uses the anti-kt algorithm, with topo-cluster inputs opo-clusters: topological groups of noise-suppressed calorimeter cells opo-clusters can be EM-scale or LCW-scale (hadronic scale) Depending on physics intent, different types of s are useful Primarily small-r (R =.4) EM s and large-r (R =.) LCW s Jets can have different underlying sources tagging Small-R mostly for quarks/gluons, large-r for energetic W/Z/H/top/X tan θ sin φ ALAS simulation E [MeV] Pythia 6.45 di event Steven Schramm (Universite de Gene ve) Jet reconstruction, calibration, and tagging.5 tan θ cos φ August, 7 /

3 Introduction Jet inputs Left: CONF-7-65 Right: CONF-7-6 Jets can be built from any set of 4-vectors opo-clusters, tracks, particle flow objects, truth particles,... opo-clusters can be pileup-suppressed (several techniques) SoftKiller, Voronoi ( ), Constituent Sub, Cluster Vertex Fraction,... Large improvements in low-p resolution in high lumi environments (Larger) s can even be built from (smaller) s: reclustering ake advantage of well-known and well-measured small-r s Unless mentioned, following slides use standard topo-clusters / R σ R ALAS Simulation Preliminary Pythia Di s= ev No Pileup Correction Anti-k LCW, R=.4 Jet-Area subtraction η true <.8, <µ>= Vor. Spread. Vor. Spread. + CVF 5 GeV Vor. Supp. + SK.6 max CS, R =.5 + SK.6 Fraction of Events.4..8 oy MC, X m X = 75 GeV, m B A B ( C D) = 8 GeV, m A = m C R(all parton pairs) >. (resolved) R(all parton pairs) >.4 and at least one R(pair) <. At least one R(pair) <. and exactly one R(pair) <.4 R(all parton pairs) <.4 (merged) = m D = partons: A, C, D for simplicity, all particles are scalars..6. Ratio to Uncorrected true p [GeV] px [GeV] Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 /

4 Calibration and uncertainties Jet calibration overview Jets need to be calibrated to account for several effects Pileup, non-compensating calorimeter response, data/mc diffs, etc he calibration chain is different for small-r and large-r s In particular, Large-R s include mass calibrations, not just energy Calibrations are becoming more similar as large-r usage grows First results of large-r in situ calibrations will be presented EM-scale small-r s (R=.4) Origin correction Pileup correction MC-based energy calibration Global sequential calibration Residual in situ energy calibration Fully calibrated small-r LCW-scale large-r s (R=.) Jet grooming (trimming) MC-based energy calibration MC-based mass calibration First results Standard approach Residual in situ energy calibration Residual in situ mass calibration Residual in situ mass calibration Fully calibrated large-r Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 4 /

5 Calibration and uncertainties Small-R pileup suppression Left: JEM-7-6 Right: PERF-6-4 Suppress pileup s by exploiting tracking information Jet Vertex agger (JV): veto if most tracks from additional vertices Particle flow: build s from a mix of track and cluster information Dynamically remove pileup energy from hard-scatter s p corr = p ρa α(n PV ) βµ Fake Jets /. / Event µ ~ p s = ev > GeV No tracker EM+JES Jets EM+JES Jets JV>.59 Particle Flow Jets Particle Flow Jets JV>. ALAS Simulation Preliminary No tracker η [GeV] p / N PV.8 ALAS Simulation.6 s = ev, Pythia Di anti-k t R =.4, EM scale Before any correction After area-based correction After residual corrections Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 5 / η

6 Calibration and uncertainties Large-R pileup suppression op: PERF-- Bottom: PERF-5- Larger radius means more susceptible to pileup Significant impact on key variables, even at high p ALAS uses trimming to counteract pileup effects Build R =. sub-s Check if pr=. < 5% p R=. Remove such sub-s Resulting s are very stable with respect to pileup Even stable for µ = [GeV] M ALAS Simulation s=8 ev ruth η <. ruth 5 < p < 5 GeV R=. s anti k t rimmed (f =5%,R =.) cut sub W s Ungroomed Slope=.9 GeV/vertex Multis (leading ) Ungroomed Slope=.7 GeV/vertex W s Groomed Slope=. GeV/vertex Multis (leading ) Groomed Slope=. GeV/vertex Number of Reconstructed Primary Vertices Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 6 /

7 Calibration and uncertainties Monte Carlo (MC) calibration Left: PERF-6-4 Right: CONF-5-7 A substantial fraction of hadronic shower energy is not measured Nuclear binding energy from strong interactions, inactive material, etc MC information is used to correct for such effects Response: X reco /X true, numerical inversion gives calibration factor After energy calibration, average is at the truth scale Small-R s: further energy corrections to fix residual effects (GSC) Shower development, flavour differences (q vs g), punch-through Large-R s: subsequent mass calibration Energy Response truth E = GeV truth E = 6 GeV truth E = GeV truth E = 4 GeV truth E = GeV ALAS Simulation s = ev, Pythia Di anti-k t R =.4, EM scale η Jet η det det Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 7 / Mass response before calibration E = GeV E = 4 GeV E = 6 GeV E = 8 GeV E = GeV E = 5 GeV ALAS Simulation Preliminary anti-k t LCW s with R =. rimmed (f =.5, R sub =.) cut Pythia 8 dis, s = 8 ev

8 Calibration and uncertainties In situ corrections Relative in situ correction Left: JEM-7- Right: CONF-7-6 Previous calibration steps assume that MC perfectly represents data It is important to fix residual data/mc differences Small-R s use η-intercalibration to do this Correct scale of forward s with respect to well-known central s Do in both data and MC to fix data/mc differences Large-R s: first studies demonstrate η dependence Balance of photon and varies, requires a correction for η. Relative response (/c) MC / data. ALAS Preliminary anti-k t R =.4, EM+JES - s = ev, L = 4.8 fb avg 85 p < 5 GeV..9 Data 6 Powheg+Pythia8 Sherpa η det ref / p leading p = B Data / MC Data 6 Pythia8 Sherpa. Stat. ALAS Preliminary - s= ev, fb γ- Events, Data 6 anti-k t R =., LC+JES rimmed ref 45 p < GeV leading η Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 8 /

9 Calibration and uncertainties In situ corrections Absolute in situ correction Left: JEM-7- Right: CONF-7-6 MC /Response Data Response Relative scale of s across the detector is now fixed he absolute scale may still differ between data and MC Investigate with three methods, using well-known reference objects Balance of Z ll and a, for low p s Balance of a photon γ and a, for medium p s Multi- balance, using a recoil system of Z/γ calibrated s he results are then combined, providing a final calibration..9 R=.4, EM+JES anti k t Data 6 γ+ Z+ Multi ALAS Preliminary s = ev, 7 fb η <.8 otal uncertainty Statistical component p [GeV] p [GeV] Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 9 / p /Response p Response MC Data.4 anti-k ALAS Preliminary t R=., LC+JES - Data 6. s = ev, - fb η <., rimmed s otal uncertainty Statistical component γ+ Multi-

10 Calibration and uncertainties In situ corrections Small-R uncertainties Both: PERF-6-4 he in situ corrections (relative and absolute) correct data to MC his correction has uncertainties, which are evaluated here are additional uncertainties beyond the in situ Flavour dependence, pileup, and punch-through are also considered -% uncertainty on the energy scale from p of 5 GeV to ev Fractional JES uncertainty Data 5, s = ev ALAS anti k t R =.4, EM+JES + in situ η =. otal uncertainty Absolute in situ JES Relative in situ JES Flav. composition, inclusive s Flav. response, inclusive s Pile up, average 5 conditions Punch through, average 5 conditions Fractional JES uncertainty Data 5, s = ev ALAS anti k t R =.4, EM+JES + in situ p = 8 GeV otal uncertainty Absolute in situ JES Relative in situ JES Flav. composition, inclusive s Flav. response, inclusive s Pile up, average 5 conditions Punch through, average 5 conditions.. 4 p [GeV] 4 4 η Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 /

11 Calibration and uncertainties In situ corrections R trk and large-r uncertainties Both: CONF-7-6 First look at large-r in situ corrections was shown However, do not yet cover full kinematic range (lack of η correction) Different procedure for uncertainty on X : R trk = (X calo/x track ) data (X calo /X track ) MC Does not correct data/mc differences, only an uncertainty on them Baseline: raw R trk difference from between data and Pythia8 Modelling: Pythia8 vs Herwig++ difference in R trk racking: tracking efficiency, momentum, and fake tracks Larger uncertainties, but quick and can be used for any variable p [GeV] Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 / Fractional JPtS uncertainty Data 6, s = ev anti-k t R =. s, LC+JES+JMS rimmed (f =.5, R sub =.) cut η <, m /p =. otal uncertainty Baseline racking ALAS Preliminary Modelling Statistical

12 Calibration and uncertainties In situ corrections Jet mass and forward folding Both: CONF-7-6 Forward folding allows for measuring the mass scale and resolution Idea: construct high-purity W or top selection, look at data and MC Shift and smear the distribution until data and MC agree his gives the mass scale and resolution data/mc difference Forward folding is very limited in where it can be derived (statistics) R trk is used to extrapolate to other kinematic regimes Events / 5 GeV ALAS Preliminary - s = ev, fb rimmed anti-k t R=. LC+JES+JMS µ + s, η <. 5 GeV < p GeV b- tag: R(b; ) < 6 Data tt Single top W+s Z+s Dibosons MC syst. error m MC m /Response data Response. ALAS Preliminary anti-k t R=., LC+JES+JMS Data 6 - s = ev, fb. η <., rimmed s <m < GeV.9 Data / MC.5.5 Jet mass [GeV] Jet mass [GeV].8.7 otal uncertainty Statistical component R trk Forward Folding p [GeV] Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 /

13 Jet substructure and tagging Hadronic object tagging Right: PERF-- Hadronic decays of massive particles can come from many sources Main examples: W qq, Z q q, H b b, top bw bqq Higher parent particle energy = decay product collimation Idea of boosted decays, rule-of-thumb R m X /p X Results in all decay products within a single Jet substructure distinguishes different parent particles (or q vs g) Expected number of decay axes W Z H x x x gluon quark (not top) top bosons? p W [GeV] Expected mass [GeV] Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 / R(q,q) ALAS Simulation Pythia Z' tt, t Wb

14 Jet substructure and tagging Jet substructure Combined mass Left: CONF-6-5 Right: JEM-7- Jet mass is the most intuitive substructure variable Signal s should have the mass of the parent particle Mass defined by energy and angular separation between constituents At very high p, calorimeter may not resolve the constituents rack-assisted mass addresses this, m A Combination m comb Fraction / 4 GeV ALAS Simulation Preliminary s = ev, W/Z-s.6 ev < p <.8 ev, η < = A(p ) m calo + B(p ) m A Uncalibrated Calibrated m calo m track m A Jet mass [GeV] Fractional mass resolution = mtrack p calo p track. ALAS Simulation Preliminary s = ev, WZ qqqq anti k t R =. s, η <. rimmed (f =.5, R =.) cut sub LCW + JES + JMS calibrated is best of both worlds Calorimeter mass rack assisted mass Combined mass ruth p Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 4 / [GeV]

15 Jet substructure and tagging Jet substructure Jet substructure Left: CONF-7-64 Right: PERF-5- Combined mass provides good separation between QCD, W, and top A good first step to tagging a hadronic decay Other substructure variables quantify further differences D β= is the most powerful variable (after mass) for W/Z tagging τ wta is the most powerful variable (after mass) for top tagging Both quantify consistency of angular distribution with signal Normalized amplitude / [GeV].8 ALAS Simulation Preliminary.6 W Jets Dis s = ev anti k t R=. s op Jets rimmed (f =.5 R sub =.) cut.4 truth p = [, 5], η truth < Combined Mass [GeV] (β=) D Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 5 / Normalised Entries. ALAS Simulation s=8 ev ruth η <. ruth 5 < p < 5 GeV M Cut anti k R=. s t rimmed (f =5%,R =.) cut sub W s (in W WZ) Multis (leading ) G & W = 5%

16 Jet substructure and tagging Jet tagging W-boson and top-quark tagging Both: CONF-7-64 W and top taggers are optimized with two-variable cuts Plots show QCD rejection as a function of p W-tagging is for a 5% signal efficiency, top uses 8% signal efficiency Boosted Decision rees and Deep Neural Networks also studied Use roughly substructure variables as inputs Reasonable gains with respect to simpler methods Background rejection ( / bkg ) DNN W BD W var optimised tagger ALAS Simulation Preliminary s = ev anti k R=. s, η truth t <. rimmed (f =.5 R sub =.) cut W tagging at sig = 5% Background rejection ( / bkg ) DNN top BD top var optimised tagger Shower Deconstruction ALAS Simulation Preliminary s = ev anti k R=. s, η truth t <. rimmed (f =.5 R sub =.) cut op tagging at sig = 8% ruth p [GeV] ruth p [GeV] Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 6 /

17 Jet substructure and tagging Jet tagging W-bosons and top-quarks in data Both: CONF-7-64 BD and DNN discriminants agree nicely between data and MC Shown for W-tagging using a DNN Application of taggers can select high purity regions Shown for top-tagging using a two-variable tagger Events /.5 Data/Pred ALAS Preliminary s = ev, 6. fb R(large, b ) >. Jet p > GeV comb Jet m > 4 GeV Data 5+6 tt (top) tt (W ) tt (other) Single op (W ) Single op (other) W +s VV, Z+s, QCD otal uncert. (no large R) Stat. uncert. tt modelling uncert rimmed large R DNN W tag discriminant Events / 5 GeV Data/Pred. Data 5+6 tt (top) tt (W ) 8 tt (other) Single op (W ) 6 Single op (other) 4 W +s VV, Z +s, QCD otal uncert. Stat. uncert. tt modelling uncert ALAS Preliminary - s = ev, 6. fb R(large, b-) <. Jet p > 5 GeV τ + d ( sig = 8%) op tag pass rimmed large-r combined mass [GeV] Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 7 /

18 Jet substructure and tagging Jet tagging H bb tagging Left: PUB-7- Right: CONF-6-9 Higgs- efficiency Hadronic top rejection ALAS Simulation Preliminary calo MVc b-tagging at 77% WP b-tag, Loose m window calo b-tags, No m selection calo b-tags, Loose m window calo b-tags, ight m window, D p [GeV] ALAS Simulation Preliminary calo MVc b-tagging at 77% WP b-tag, Loose m window calo b-tags, No m selection calo b-tags, Loose m window calo b-tags, ight m window, D p [GeV] sel. sel. Double Sub B Labelling Efficiency H-tagging uses b-tagging and substructure information b-tagging: QCD rejection Substructure: top, g bb At high p, track-s for b-tagging merge (truth-level) New types recover eff. ALAS Simulation Preliminary 76 GeV < m R=. rack Jet VR rack Jet ExKt Sub CoM Sub Higgs Jet p Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 8 / < 46 GeV [GeV]

19 Jet substructure and tagging Jet tagging Quark vs gluon tagging Left: PUB-7-9 Right: PUB-7-7 Differentiating quarks from gluons is different in several ways Jet mass is not useful (not a hadronically decaying massive particle) Uses small-r s rather than large-r s Best single variable is track multiplicity (gluons radiate more) Also tried a convolutional neural network (CNN) Some gains over a two-variable tagger for high quark efficiencies Normalized Entries ALAS Simulation Preliminary s = ev Anti k EM+JES R=.4 <. Quark Jet Gluon Jet 5 < p < GeV 4 < p < 5 GeV < p < 5 GeV n track Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 9 /

20 Jet substructure and tagging Future prospects agging into the future Both: PUB-7-5 Fractional mass resolution Calorimeter s ability to resolve highly boosted decays is degraded Ideal: angular resolution of the tracker and scale of the calorimeter ALAS is developing a substructure-oriented particle flow CC Combined mass currently superior at low p, CC better at high p CC vastly superior for non-mass substructure variables CC algorithm promises large gains to future hadronic object tagging ALAS Simulation Preliminary s= ev anti k R=., WZ qqqq η <., p > GeV LC opo (trimmed + calibrated + m ) comb All CCs (trimmed + not calibrated) Combined CCs (not trimmed + not calibrated) ruth p [GeV] resolution Jet D ALAS Simulation Preliminary s= ev anti k R=., WZ qqqq η <., p > GeV LC opo (trimmed + calibrated + m All CCs (trimmed + not calibrated) ) comb Combined CCs (not trimmed + not calibrated) ruth p Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 / [GeV]

21 Summary Summary ALAS s are typically built of calorimeter topo-clusters Jets built using particle flow objects are also starting to be used ALAS makes use of small-r (.4) and large-r (.) s Both types are calibrated in several steps Account for pileup, calorimeter response, and data/mc differences Jets are also used to tag hadronic showers W, Z, H, and top tagging using large-r s quark vs gluon discrimination using small-r s Machine learning methods (BDs, DNNs, CNNs) studied for tagging Work ongoing to improve highly boosted substructure and tagging his talk focused on current techniques, used in 5 and 6 Lots of work is ongoing to prepare for the future! Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 /

22 Backup Backup Material Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 /

23 Backup Constituent-level pileup suppression All: CONF-7-65 / R σ R Ratio to Uncorrected ALAS Simulation Preliminary Pythia Di s= ev No Pileup Correction Anti-k LCW, R=.4 Jet-Area subtraction η true <.8, <µ>= Vor. Spread. Vor. Spread. + CVF 5-GeV Vor. Supp. + SK.6 max CS, R =.5 + SK true p [GeV] Small gains compared to nominal ( areas) in low pileup environment Large gains by µ = / R σ R ALAS Simulation Preliminary Pythia Di s= ev No Pileup Correction Anti-k LCW, R=.4 Jet-Area subtraction η true <.8, <µ>=8 Vor. Spread. Vor. Spread. + CVF 5-GeV Vor. Supp. + SK.6 max CS, R =.5 + SK.6 / R σ R ALAS Simulation Preliminary Pythia Di s= ev No Pileup Correction Anti-k LCW, R=.4 Jet-Area subtraction η true <.8, <µ>= Vor. Spread. Vor. Spread. + CVF 5 GeV Vor. Supp. + SK.6 max CS, R =.5 + SK Ratio to Uncorrected true p [GeV] Ratio to Uncorrected true p [GeV] Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 /

24 Backup rimming at high luminosity All: JetSubstructureECFA4 Normalized entries.6 ALAS Simulation Preliminary LCW s with R=. anti k t.4.. No grooming, no pileup correction s = 4 ev, 5 ns bunch spacing η <., 5 < p < GeV Pythia8 dis µ= µ=8 µ=4 µ= µ= Normalized entries.5 ALAS Simulation Preliminary anti k t LCW s with R=. No grooming, no pileup correction. s = 4 ev, 5 ns bunch spacing.5 η <., 5 < p < GeV Pythia8 Z tt (m = ev) Z µ= µ=8 µ=4 µ= µ= Leading mass [GeV] Leading mass [GeV] Normalized entries.4 ALAS Simulation Preliminary. anti k t LCW s with R=. µ=. rimmed, pileup corrected µ=8.8 s = 4 ev, 5 ns bunch spacing µ=4 η <., 5 < p < GeV µ=.6 Pythia8 dis µ= Normalized entries.5 ALAS Simulation Preliminary anti k t LCW s with R=. rimmed, pileup corrected. s = 4 ev, 5 ns bunch spacing.5..5 η <., 5 < p < GeV Pythia8 Z tt (m = ev) Z µ= µ=8 µ=4 µ= µ= Leading mass [GeV] Leading mass [GeV] Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 4 /

25 Backup Jet substructure uncertainties Both: JEM-6-9 he R trk method is used to derive substructure uncertainties Shown for D β= (W/Z tagging) and τ wta (top tagging) Plots are older (5 data only), and thus are statistically limited Update with 6 data reduces the statistical component significantly Uncertainties are under control, so they can be used for object tagging Fractional JD S uncertainty. Data 5, s = ev ALAS Preliminary anti-k t R =. s, LCW+JES+JMS.5 rimmed (f =.5, R sub =.), p > 5 GeV cut η <, m /p = otal uncertainty R trk baseline (Pythia 8) R trk modelling (Herwig++) R trk tracking Statistical wta S uncertainty Fractional Jτ. Data 5, s = ev ALAS Preliminary anti-k t R =. s, LCW+JES+JMS.5 rimmed (f =.5, R sub =.), p > 5 GeV cut η <, m /p = otal uncertainty R trk baseline (Pythia 8) R trk modelling (Herwig++) R trk tracking Statistical p [GeV] p [GeV] Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 5 /

26 Backup Jet reclustering Both: CONF-7-6 Idea: small-r s are already carefully calibrated and understood Build large-r s from them, take advantage of small-r knowledge Allows for choosing the radius you want, not only R =. ypically smaller uncertainties in the range of interest (p ev) Also improved mass resolution, compared to m calo Performance degrades at higher p (increased boost) Effective reclustered size becomes one small-r (not m comb ) Jet Energy Scale (JES) Uncertainty [%] 5 4 ALAS Preliminary s W' = ev WZ, m > 5 GeV Reclustered (R=., f =5%) from cut anti-k t R=.4 EM+JES+GSC+JMS s rimmed (f =5%, R =.) cut sub anti-k t R=. LC+JES+JMS s Reclustered (from small-r JES) rimmed rimmed (f =5%, R =.) anti-k R=. LC+JES+JMS s. cut sub t Jet p [GeV] true p [GeV] Steven Schramm (Université de Genève) Jet reconstruction, calibration, and tagging August, 7 6 / Fractional mass resolution ALAS Simulation Preliminary Pythia 8 W' WZ; W qq; η <.4 Calorimeter-only mass Reclustered (R=., f =5%) from anti-k cut t R=.4 EM+JES+GSC+JMS s